On Sat, Feb 28, 2026 at 09:33:42PM +0100, Sebastian Ramacher wrote:
> On 2026-02-24 09:58:15 +0100, Dylan Aïssi wrote:
> > Hi,
> >
> > On 29/01/2026 20:45, Adrian Bunk wrote:
> > > > Oups, actually tests segfault during the initiation step (so before
> > > > testing onnxruntime), this seems to be an issue with either ONNX or
> > > > PyTorch.
> > > > I proposed a new autopkgtest to onnx to catch this kind of bugs earlier:
> > > > https://salsa.debian.org/deeplearning-team/onnx/-/merge_requests/5
> > >
> > > Your manual test from #1126427
> > > import torch
> > > import onnx
> > > fails for me in unstable but works after downgrading to
> > > python3-onnx/forky.
> > >
> > > IMHO onnx should not migrate to testing when it isn't understood what is
> > > going wrong there.
> > >
> > Fwiw, building ONNX with ONNX_USE_LITE_PROTO set to ON prevents the crash:
> > https://salsa.debian.org/deeplearning-team/onnx/-/merge_requests/6
> >
> > But, it seems that we need to rebuild at least onnxruntime (and pytorch?)
> > against this new ONNX.
>
> Scheduled a rebuild of onnxruntime.
This had no effect due to ONNX not yet uploaded.
The good news is that changing ONNX and then rebuilding onnxruntime
fixes my minimal testcase in #1126427.
The bad news is that "we need to rebuild at least" is because the ONNX
change changes the ABI:
ImportError:
/usr/lib/python3/dist-packages/onnxruntime/capi/onnxruntime_pybind11_state.cpython-313-x86_64-linux-gnu.so:
undefined symbol: _ZN4onnx39AttributeProto_AttributeType_descriptorEv
This would need a package rename[1] with transition:
Package: libonnx1l
Replaces: libonnx1, libonnx1t64
Breaks: libonnx1 (<< ${source:Version}), libonnx1t64 (<< ${source:Version})
<remove time_t Provides>
rdeps in the transition:
- onnxruntime
- pytorch
- pytorch-cuda (not in testing)
- toppic
The 3 packages in testing build for me with the changed ONNX.
> Cheers
cu
Adrian
[1] there might be a better package name than appending 'l'
